Détection d'évènements à partir de Twitter

We present a system for finding, from Twitter data, events that raised the interest of users within a given time period and the important dates for each event. An event is represented by many terms whose frequency increases suddenly at one or more moments during the analysed period. In order to determine the terms (especially the hashtags) dealing with a topic, we propose methods to cluster similar terms: phonetic methods adapted to the writing mode used by users and some statistical methods. In order to select the set of events, we used three main criteria: frequency, variation and Tf·Idf. MOTS-CLÉS : Twitter, hashtags, évènement, similarité sémantique, DBscan.